Automatic Building Footprint Generation from Airborne Lidar Point Cloud

Automatically generating building footprint from remote sensing data is an active research topic because of the widespread usage of building footprint in numerous applications. The invention of airborne LiDAR technology has made it possible to measure the ground objects in a large-scale area with a dense and accurate three-dimensional (3-D) point cloud, and therefore provide a new and promising data source for extracting building footprint. However, due to the fact that the LiDAR point cloud is a set of unordered 3-D point coordinates with tremendous size, many traditional remote sensing algorithms that are designed for processing raster and image data cannot be directly applied on LiDAR point cloud. This research presents an efficient and automated workflow to generate building footprint from pre-classified LiDAR data. In this workflow, the pre-classified LiDAR points that belong to the building category are first segmented into multiple clusters through applying an efficient grid-based segmentation algorithm. Each cluster contains the points of an individual building. Then the recursive convex hull algorithm is designed and applied on each cluster to efficiently generate the initial outline for each building. The LiDAR points are irregularly distributed, which causes the generated vii initial building outline to contained irregular zig-zag shape. The initial building outline needs to be regularized in order to deliver the final building footprint with acceptable linear or curvilinear boundaries. To achieve this, a signal-based regularization algorithm that can analyze the wholistic geometric structure of building outline through a 1-D signal is introduced. The signalbased regularization uses Gaussian Smoothing and unsupervised data clustering as the main techniques to regularize the initial building outline. In order to improve it, the more advanced signal processing technique named Cauchy Norm Decomposition is also proposed for more effective regularization. Furthermore, for the purpose of generating final building footprint for the building that may have curvilinear boundary, a robust regularization algorithm that is able to reconstruct both straight-line and curvilinear boundaries is developed by denoising the cumulative signal transformed from initial building outline. The performance of grid-based segmentation and recursive hull algorithm are evaluated qualitatively using the datasets collected at both Santa Rosa, CA and Toronto Downtown. The performance of all the regularization algorithm is evaluated qualitatively and quantitatively using the same datasets.